AI Music: The Tech and Business of Suno / Udio / MusicGen

L5-05 covered TTS (making AI talk). This piece: making AI write complete songs — lyrics + vocals + instruments + arrangement.

Early 2024 AI music still “sounded like AI.” By 2025-2026, Suno v5 + Udio v3 are indistinguishable from human-made tracks in many cases for average listeners.

The big three

1. Suno (2023+)

The most popular AI music generator:

• Position: consumer-grade + easy + fast
• v5 released 2025-09: major leap in vocal clarity, genre accuracy, production quality
• Features: Song Editor (section-level Replace / Extend / Crop), Studio mode (DAW-style + 12-track stem separation)
• Pricing: Basic free / Pro $10/mo / Premier$30/mo (includes commercial use + Studio)

2. Udio (2024+)

A competitor focused on quality over speed:

• Position: pristine lead vocals + cohesive harmonies
• Features: transparent credit accounting, excellent vocal clarity
• Pricing: Free $0 / Standard ~$10 / Pro ~$30

Suno vs Udio: Suno has stronger editor + DAW tools; Udio has cleaner vocals + better harmonies.

3. MusicGen (Meta open-source, 2023)

The open-source camp:

• Weights public (HuggingFace)
• Quality below Suno / Udio but good for research
• Best for local deployment / custom development

Architecture (the public parts)

Suno / Udio haven’t published their full training pipelines, but the community consensus components:

[Text prompt + style description]
    ↓
Text encoder (CLIP / T5-style)
    ↓
Music tokenizer (audio → discrete token sequence)
    ↓
Transformer / Diffusion main model
    ↓
Music tokens
    ↓
Audio decoder (vocoder, e.g. EnCodec / SoundStream)
    ↓
[Final waveform]

Two main approaches:

A. Autoregressive (e.g. MusicGen / early Suno)

• Audio compressed to discrete tokens (like NLP’s BPE)
• Transformer generates token by token
• Pros: controllable; Cons: slow generation

B. Diffusion (Stable Audio / Riffusion)

• Diffusion in latent audio space
• Pros: high quality, fast; Cons: less controllable

Suno v5 / Udio v3: industry guesses a hybrid of both.

Evaluation challenges

Audio quality is hard to evaluate automatically:

• FAD (Fréchet Audio Distance): like FID but for audio features
• MOS (Mean Opinion Score): human 1-5 ratings (gold standard)
• Real vs AI discrimination: can listeners tell?

In 2024 blind tests, Suno v4 already fooled half the listeners into thinking it was human-made; v5 + Udio closed the gap further.

Legal / copyright fights

AI music is the fiercest copyright battlefield of 2024-2026:

1. Record label lawsuits

2024-06: RIAA, representing Universal / Sony / Warner, sued Suno and Udio:

• Allegation: training data includes copyrighted music
• Crux: does “Fair Use” apply to training data?
• Status: ongoing

2. Style copying vs output copying

Two different problems:

• Training-data copying: the model has seen Taylor Swift’s songs
• Output copying: the model generates songs that sound like Taylor Swift

The first is a “training-stage” problem; the second is an “output-stage” problem — Suno / Udio both added output filters to avoid direct mimicry.

3. Artist activism

2024-04: 200+ artists (Billie Eilish, Stevie Wonder, Nicki Minaj, et al) signed an open letter opposing “unauthorized AI training.”

4. Commercial use

• Suno Pro / Premier provides “commercial license” — but its legal standing is untested
• Using AI music in YouTube videos may trigger Content ID flags
• Labels may pursue royalties retroactively

Use cases

Already used:

• YouTube / TikTok BGM (individual creators)
• Ads / marketing assets (SMBs)
• Podcast intros
• Game / app sound effects (especially budget indie games)

Advanced:

• Pro production assist: AI generates drafts → human rearranges
• Personalized birthday / wedding songs (consumer product)
• Music education: generate examples in different styles

Probably won’t be replaced:

• Top-tier original artists
• Live performance
• Emotionally narrative albums

Pricing comparison (2026)

Tool	Free tier	Mid tier	Top tier	Commercial
Suno	50 credits/day	$10/mo (Pro)	$30/mo (Premier)	Pro+
Udio	10 credits/day	$10/mo (Standard)	$30/mo (Pro)	Pro
MusicGen	Open source	-	-	Check license
Stable Audio	Limited free	~$12/mo	~$30/mo	Paid tiers

$10/mo gets you hundreds of songs — why independent creators have adopted it en masse.

Practical prompt tips

Tip	Example
Specify genre	”lo-fi hip-hop, jazzy piano, mellow” beats “good music” 10×
Specify BPM	”120 BPM, drum and bass”
Specify structure	”verse, chorus, bridge, outro”
Constraints / excludes	”no vocals, instrumental only”
Style reference	”in the style of 90s Britpop”
Mood + use case	”uplifting, for product launch video”

Suno’s lyric prompt syntax (special):

[Verse 1]
... lyrics ...

[Chorus]
... chorus ...

[Bridge]
... bridge ...

[Instrumental Solo]

Bracket tags control structure.

Industry impact

Short term:

• Mid-to-low-end commercial music market (background / stock music) under attack
• Suno / Udio already poaching customers from stock music companies (Epidemic Sound et al)

Mid term:

• Musicians start using AI for drafts / arrangement / instrumentation
• Deeper DAW (Logic / Ableton) integration

Long term:

• Personalized music (“your unique BGM”) becomes cheap
• Boundary between “music” and “content” blurs (dynamic OSTs)

Why music AI lagged image AI by 2 years

1. Data scarcer: high-quality audio data smaller than image data
2. Time dimension: music has temporal structure / consistency requirements
3. Aural sensitivity: human ear is sensitive to artifacts
4. Tighter copyright: music industry has mature IP chains that fight fast

But catching up extremely fast — Stable Diffusion 2022 → Suno v5 in only 3 years.

Recommended further reading

• HelloAI L5-05 TTS
• HelloAI L5-02 Diffusion Math
• HelloAI L6-05 Bias & Fairness (copyright / cultural issues in AI music)
• Suno official
• Udio official
• MusicGen paper

🚧 3 Common Pitfalls